ESE PhD candidate Nikolaos Bousias defended his thesis on Symmetries Infused Safe & Scalable Multi-Robot Policies on March 11th. Advised by Professor George Pappas, Bousias’ thesis focused on overcoming challenges in ML for multi-robot systems.
“We live in an era where autonomous agents are no longer a futuristic idea. In many applications, a team of robots can do more: they can inspect larger areas, complete tasks faster, and provide redundancy and robustness. But this also introduces new challenges, because now we need not only autonomy at the individual level, but also coordination at the team level,” Bousias explains. According to Bousias, safety and scalability remains the primary challenge in multi-robot systems. With more agents introduced, the complexity of our algorithms soon becomes computationally intractable, posing fundamental challenges for traditional optimization and machine learning approaches.
Instead of relying on pure black-box AI, leveraging the symmetries, exhibited in physical structures, allows engineers to curtail the problem into a smaller, more manageable one, Bousias explains.
“We can take those symmetries and embed them into our neural networks as an inductive bias, so that we get enhanced policy generalization, sample efficiency, safety and scalability.” Doing so allows AI models to remain “structurally faithful to the actual physical systems that we’re trying to learn a policy or safety-certificates for.”
For Bousias, the challenging aspects of his research resulted in problem identification. “The hardest part of engineering isn’t finding a solution but identifying the problem.” According to Bousias, sometimes the right problem “is to figure out a problem that is not a problem today, but it’s going to be a problem tomorrow.”
Bousias attributes his ability to overcoming research challenges with the mental toughness he acquired from his years in competitive sailing in Greece. “You learn that you have more to give. You go for it. You push even further.” Those formative experiences with sailing gave Bousias insight into how to deal with failure, lessons he applied during his PhD research. “You will fall down seven times, so make sure you get up the eighth,” Bousias says. Healthy skepticism of one’s research remains essential as well. “Science only moves forward when researchers constantly question their own results and absolutes.”
The collaborative community at Penn remained instrumental to his success. In fact, Bousias attributes the inspiration for his thesis research to collaborations with fellow doctoral student colleagues such as Mariliza Tzes, Stefanos Pertigkiozoglou, and Evangelos Chatzipantazis. Winning Best Paper for Graph Neural Networks for Multi-Robot Active Information Acquisition at the International Conference on Robotics and Automation (ICRA 2023) helped propel him forward. According to Bousias, while the award was “quite the high,” it wasn’t the prestige that mattered as much as the sense of community it engendered. “It helps re-define your path forward because you’re getting feedback from all of these people. And suddenly you start thinking in a different way,” Bousias says. Feedback from his advisor Professor George Pappas, as well as committee members Professor Alejandro Ribeiro, Professor Nikolai Matni, and Professor Amanda Prorok also played a pivotal role. So did support from his family, particularly his sister who joked he was the “last person in the family to become a doctor, and not even the right type of doctor.”
Regarding future plans, Bousias is interviewing for positions in both industry and postdoctoral appointments in academia.
When he’s not working, Bousias is an avid reader of history and philosophy. In addition, he loves windsurfing and cinematography, particularly the works of Paolo Sorrentino and Ernst Bergman, which often inspire the screenings and discussions at the film club he started, Φ.Α.Κ.Α..
Learn more about Bousias’ work here